Predicting isocitrate dehydrogenase status among adult patients with diffuse glioma using patient characteristics, radiomic features, and magnetic resonance imaging: Multi-modal analysis by variable vision transformer

• Published in: Magnetic resonance imaging Vol. 111; pp. 266 - 276
• Main Authors: Usuzaki, Takuma, Inamori, Ryusei, Shizukuishi, Takashi, Morishita, Yohei, Takagi, Hidenobu, Ishikuro, Mami, Obara, Taku, Takase, Kei
• Format: Journal Article
• Language: English
• Published: Netherlands Elsevier Inc 01.09.2024
• Subjects: Artificial intelligence; Brain tumor; Deep learning; Genetic mutation; Neural networks; Brain tumor; Deep learning; Neural networks; Artificial intelligence; Genetic mutation
• ISSN: 0730-725X; 1873-5894; 1873-5894
• DOI: 10.1016/j.mri.2024.05.012

To evaluate the performance of the multimodal model, termed variable Vision Transformer (vViT), in the task of predicting isocitrate dehydrogenase (IDH) status among adult patients with diffuse glioma. vViT was designed to predict IDH status using patient characteristics (sex and age), radiomic features, and contrast-enhanced T1-weighted images (CE-T1WI). Radiomic features were extracted from each enhancing tumor (ET), necrotic tumor core (NCR), and peritumoral edematous/infiltrated tissue (ED). CE-T1WI were split into four images and input to vViT. In the training, internal test, and external test, 271 patients with 1070 images (535 IDH wildtype, 535 IDH mutant), 35 patients with 194 images (97 IDH wildtype, 97 IDH mutant), and 291 patients with 872 images (436 IDH wildtype, 436 IDH mutant) were analyzed, respectively. Metrics including accuracy and AUC-ROC were calculated for the internal and external test datasets. Permutation importance analysis combined with the Mann–Whitney U test was performed to compare inputs. For the internal test dataset, vViT correctly predicted IDH status for all patients. For the external test dataset, an accuracy of 0.935 (95% confidence interval; 0.913–0.945) and AUC-ROC of 0.887 (0.798–0.956) were obtained. For both internal and external test datasets, CE-T1WI ET radiomic features and patient characteristics had higher importance than other inputs (p < 0.05). The vViT has the potential to be a competent model in predicting IDH status among adult patients with diffuse glioma. Our results indicate that age, sex, and CE-T1WI ET radiomic features have key information in estimating IDH status. •Multimodal analysis using demographics, radiomic features, and MRI was performed by vViT.•The accuracy vViT achieved for internal and external datasets was 1.00 and 0.935, respectively.•Our analyses revealed that the radiomic features of the contrast-enhanced part had key information for IDH prediction.